GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.
Preprint
10 Pith papers cite this work, alongside 7 external citations. Polarity classification is still indexing.
10
Pith papers citing it
7
external citations · Pith
abstract
Understanding Transformer-based models has attracted significant attention, as they lie at the heart of recent technological advances across machine learning. While most interpretability methods rely on running models over inputs, recent work has shown that a zero-pass approach, where parameters are interpreted directly without a forward/backward pass is feasible for some Transformer parameters, and for two-layer attention networks. In this work, we present a theoretical analysis where all parameters of a trained Transformer are interpreted by projecting them into the embedding space, that is, the space of vocabulary items they operate on. We derive a simple theoretical framework to support our arguments and provide ample evidence for its validity. First, an empirical analysis showing that parameters of both pretrained and fine-tuned models can be interpreted in embedding space. Second, we present two applications of our framework: (a) aligning the parameters of different models that share a vocabulary, and (b) constructing a classifier without training by ``translating'' the parameters of a fine-tuned classifier to parameters of a different model that was only pretrained. Overall, our findings open the door to interpretation methods that, at least in part, abstract away from model specifics and operate in the embedding space only.
representative citing papers
The paper introduces compositional interpretability as a category-theoretic framework that casts mechanistic explanations as commuting syntactic-semantic mappings optimized under faithfulness and complexity constraints derived from minimum description length.
Gated SAEs decouple which features to use from how large their activations should be, applying the L1 penalty only to selection and thereby eliminating shrinkage while halving the number of firing features needed for good fidelity.
Training per-layer affine probes on frozen transformers yields more reliable latent predictions than the logit lens and enables detection of malicious inputs from prediction trajectories.
MEMIT scales direct memory editing in transformers from single facts to thousands of associations by optimizing MLP weight updates.
A lightweight Q-Former proxy trained on VLM hidden states reveals that localization signals peak in input-dependent intermediate layers, not the final layers used by standard editing pipelines.
A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.
Language models contain localized entity-selective neurons in early layers that causally mediate factual recall for specific entities across surface variations.
Introduces effective dimension d_ρ from spectral analysis of reasoning trajectories to distinguish task hardness (0.93 AUC on MATH500) and uses kinematic features for early correctness prediction from partial generations.
Perceptual geometry for color, pitch, emotion and taste emerges transiently in intermediate layers of transformer LLMs despite purely textual training.
citing papers explorer
-
Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small
GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.
-
From Mechanistic to Compositional Interpretability
The paper introduces compositional interpretability as a category-theoretic framework that casts mechanistic explanations as commuting syntactic-semantic mappings optimized under faithfulness and complexity constraints derived from minimum description length.
-
Improving Dictionary Learning with Gated Sparse Autoencoders
Gated SAEs decouple which features to use from how large their activations should be, applying the L1 penalty only to selection and thereby eliminating shrinkage while halving the number of firing features needed for good fidelity.
-
Eliciting Latent Predictions from Transformers with the Tuned Lens
Training per-layer affine probes on frozen transformers yields more reliable latent predictions than the logit lens and enables detection of malicious inputs from prediction trajectories.
-
Mass-Editing Memory in a Transformer
MEMIT scales direct memory editing in transformers from single facts to thousands of associations by optimizing MLP weight updates.
-
Analysis-by-Proxy: Localization Signals in VLMs Operating as Condition Encoders
A lightweight Q-Former proxy trained on VLM hidden states reveals that localization signals peak in input-dependent intermediate layers, not the final layers used by standard editing pipelines.
-
Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces
A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.
-
Friends and Grandmothers in Silico: Localizing Entity Cells in Language Models
Language models contain localized entity-selective neurons in early layers that causally mediate factual recall for specific entities across surface variations.
-
Geometric Signatures of Reasoning: A Spectral Perspective on Task Hardness
Introduces effective dimension d_ρ from spectral analysis of reasoning trajectories to distinguish task hardness (0.93 AUC on MATH500) and uses kinematic features for early correctness prediction from partial generations.
-
Geometry of Human Perceptual Domains Emerges Transiently in LLM Representations
Perceptual geometry for color, pitch, emotion and taste emerges transiently in intermediate layers of transformer LLMs despite purely textual training.